CASIA OpenIR  > 毕业生  > 博士学位论文
指纹和人脸中的特征提取和分析
其他题名Feature Extraction and Analysis of Fingerprint and Face Images
何余良
学位类型工学博士
导师田捷
2006-02-19
学位授予单位中国科学院研究生院
学位授予地点中国科学院自动化研究所
学位专业模式识别与智能系统
关键词特征间关联性表达 测地线关联特性 概率距离 映射模型 统计模型 主动形状模型和主动表观模型 Relation Representation Among Features Geodesic Feature Possibility Distance Mapping Model Statistical Model Active Shape Model (Asm) Active Appearance Model (Aam)
摘要现代信息处理技术涉及到模式识别、人工智能、计算机视觉、机器学习、心理生物学和认知等科学。而模式识别因其明确的问题定义、严格的数学基础、坚实的理论框架和广泛的应用价值而获得越来越多的重视,并且已有丰富的理论成果,并被广泛地应用到生物身份认证、图像理解、信息检索等领域。而在模式识别系统中,特征提取是系统中至关重要的环节,其实质是从原始数据中抽取能反映样本且能区分不同类别样本的特征的过程。在此过程中包括特征表达、特征筛选、特征评估等内容。 本文的主要工作集中在如下两个方面:第一,作者从统计分析角度分析现有图像识别系统中的特征提取方法和特征之间的关联性,用概率距离做为区分特征的度量准则来评估特征选择的性能。第二,作者将这种方法应用于指纹和人脸图像的特征提取和分析中。 总结起来,本文的主要贡献体现在如下三个方面: 第一、提出了一种基于特征间的二元关联特性的特征表达方法,并应用于指纹表达。在该方法中,以现有鲁棒的图像特征(包括用Fourier描述子描述图像中的曲线、用Gabor描述子描述图像的纹理等)和特征间的关联特性为基本元素,建立图像的整体特征模型来表示图像对象。在本文的第三章中,作者将这种方法应用到指纹特征的关联性分析,提出了复合细节点、二元欧式关联特性和二元测地线关联特性,并以特征点的二元结构为基础,进行指纹匹配。在第五章中,改进了基于零子空间的线性判别分析算法,其目的是从统计角度提取对象的可区分性特征。 第二、提出了一种基于k-均值聚类方法的图像映射模型的估计方法。聚类方法的目的是以同类样本特征间的差异应尽可能小而异类样本特征值间的差异应尽可能大为目标。在该方法中,作者用概率距离来衡量特征间的距离,评估特征分组性能。在第四章中,作者将该方法应用到指纹匹配过程中的指纹映射模型的估计。在指纹映射模型的估计过程中,用k-均值聚类方法将指纹匹配过程中的各个区域映射模型进行分组,分组结果用概率距离来评估,选择出最优的映射模型。 第三,介绍了一种基于统计模型进行特征提取的框架。在该框架下,用已知的特征模型和训练样本,建立可表达的、低维空间下的对象模型,并采用主动搜索方法在图像中搜索到最佳的图像特征。在第六章中,作者实现了基于统计模型的主动形状模型和主动表观模型,并应用于人脸特征点定位。
其他摘要Modern information technology includes techniques in pattern recognition, artifical intelligence, computer vision, machine learning, psychology, and so on. Among these techonologies, pattern recognition has become more and more popular for its explicit question definition, precise mathematic foundation, and rich theoretical background. It also has broad applications, including automatic biometrical authentication, image understanding and information retrieval. In a pattern recognition application system, feature extraction, as one of its key tasks, is to extract the distinguishable feature from raw data. Feature representation, filtering and evaluation need to be done in the system too. This paper includes two main works as follows: One is a stastical feature selection method, including statistical model–based feature extraction, relation representation among feature, and entropy distance-based feature evaluation. Another is its applications in fingerprint feature and facial feature analysis. In summary, the contributions of this paper are listed as follows: 1) An object representation method with binary relations among features are proposed and applied into fingerprint representation. First, Fourier descriptor-based curve representation and Gabor-wavelet descriptor-based texture representation are used as elements of an image object. Then, two kinds of binary relation among features in Euclidean space and Geodesic space are proposed to respectively connect all isolated features one another to wholly represent an image object. Finally, these schemes are applied into fingerprint representation in Chapter three, such as comprehensive minutiae, Euclidean-space binary relation and Geodesic-space binary relation among comprehensive minutiae. Additionally, improved null-space linear discriminant analsysis (NLDA) are used to extract global distinguishable features in terms of statistics. 2) A mapping model estimation method with k-means clustering is proposed in this paper in terms of possibility distance. The main target of clustering is to seek the minimum distance among homogeneous features and the maximum distance among heterogeneous features. And possibility distance is used to measure the distance between features and to evaluate grouping of features. In Chapter four, k-means clustering method is applied to classify local mapping models and their results are evaluated using possibility distance for the global optimal mapping model. 3) A framework of statistical model is introduced for feature extraction from images. With the framework, for a known high-dimension features and their training samples, a model is built by descreasing the dimensions of these features with principinal component analysis (PCA), and then an optimal feature are actively extracted from an input sample. Based on this idea, active shape model (ASM) and active appearance model (AAM) are applied to facial feature point detection in Chapter six.
馆藏号XWLW964
其他标识符200318014603008
语种中文
文献类型学位论文
条目标识符http://ir.ia.ac.cn/handle/173211/5891
专题毕业生_博士学位论文
推荐引用方式
GB/T 7714
何余良. 指纹和人脸中的特征提取和分析[D]. 中国科学院自动化研究所. 中国科学院研究生院,2006.
条目包含的文件
文件名称/大小 文献类型 版本类型 开放类型 使用许可
CASIA_20031801460300(1935KB) 暂不开放CC BY-NC-SA请求全文
个性服务
推荐该条目
保存到收藏夹
查看访问统计
导出为Endnote文件
谷歌学术
谷歌学术中相似的文章
[何余良]的文章
百度学术
百度学术中相似的文章
[何余良]的文章
必应学术
必应学术中相似的文章
[何余良]的文章
相关权益政策
暂无数据
收藏/分享
所有评论 (0)
暂无评论
 

除非特别说明,本系统中所有内容都受版权保护,并保留所有权利。